WO2000014886A1

WO2000014886A1 - Device and method for entropy encoding of information words and device and method for decoding entropy-encoded information words

Info

Publication number: WO2000014886A1
Application number: PCT/EP1999/005859
Authority: WO
Inventors: Martin Dietz; Ali Nowbakht-Irani; Ralph Sperschneider; Oliver Kunz
Original assignee: Fraunhofer-Gesellschaft zur Förderung der angewandten Forschung e.V.
Priority date: 1998-09-07
Filing date: 1999-08-12
Publication date: 2000-03-16
Also published as: US6441755B1; AU747694B2; KR20010073132A; JP2002524960A; ATE214211T1; CA2341864A1; DE59900962D1; CN1322405A; KR100397806B1; CN1185795C; DE19840835C2; AU5732999A; JP3417479B2; HK1040327A1; EP1112621A1; CA2341864C; DE19840835A1; EP1112621B1; HK1040327B

Abstract

The invention relates to a device and method for entropy encoding and associated decoding. The inventive code comprises a code table with reversible code words (12) and an escape region for information words to be coded which are located outside the region (14) defined by said code table. Said region can be selected in such a way that a major part of the information words is coded with symmetrical code words by means of the code table. On the one hand, it is possible to carry out forward decoding and also backward decoding (24) and on the other hand, use of reversible code words allows for rapid recognition of errors in a code word stream transmitted over a non-ideal channel.

Description

Device and method for entropy-coding information words and device and method for decoding entropy-coded information words

description

The present invention relates to a concept for entropy coding and a corresponding concept for decoding entropy-coded information words. In particular, the present invention relates to the error-proof entropy coding and corresponding decoding of audio signals.

Modern audio coding methods or decoding methods, which operate for example according to the MPEG layer 3 standard, are able to compress the data rate of audio signals by a factor of 12, for example, without noticeably deteriorating the quality thereof. In order to achieve such a high data rate reduction, an audio signal is sampled, whereby a sequence of discrete-time samples is obtained. As is known in the art, this sequence of discrete-time samples is windowed using suitable window functions to obtain windowed blocks of temporal samples. A block of time-windowed samples is then transformed into the frequency domain by means of a filter bank, a modified discrete cosine transform (MDCT) or other suitable device in order to obtain spectral values that collectively contain the audio signal, that is to say the temporal segment that passes through the block of time-discrete samples is given in the frequency domain. Usually 50% overlapping temporal blocks are generated and transformed into the frequency range by means of an MDCT, so that due to the special properties of the MDCT, for example, 1024 time-discrete samples lead to 1024 spectral values. It is known that the receptivity of the human ear depends on the instantaneous spectrum of the audio signal itself. This dependence is recorded in the so-called psychoacoustic model, by means of which it has long been possible to calculate masking thresholds depending on the current spectrum. Masking means that a certain tone or spectral component is masked if, for example, an adjacent spectral range has a relatively high energy. This fact of masking is used to quantify the spectral values after the transformation as roughly as possible. The aim is therefore on the one hand to avoid audible interference in the decoded audio signal and on the other hand to use as few bits as possible in order to encode or quantize the audio signal here. The interference introduced by the quantization, ie the quantization noise, should be below the masking threshold and should therefore be inaudible. According to known methods, the spectral values are therefore divided into so-called scale factor bands, which should correspond to the frequency groups of the human ear. Spectral values in a scale factor group are multiplied by a scale factor in order to scale spectral values of a scale factor band as a whole. The scale factor bands scaled by the scale factor are then quantized, whereupon quantized spectral values arise. Of course, grouping into scale factor bands is not critical. However, it is used in the standards MPEG-Layer 3 and the standard MPEG-2 AAC (AAC = Advanced Audio Coding).

A very important aspect of data reduction is the entropy coding of the quantized spectral values that follows after quantization. Huffman coding is usually used for entropy coding. Huffman coding is a coding of variable length, ie the length of the code word for a value to be coded depends on its probability of occurrence. Logically you rank the most likely Sign the shortest code, ie the shortest code word, so that a very good redundancy reduction can be achieved with the Huffman coding. An example of a well-known coding with general length is the Morse alphabet.

In audio coding, Huffman codes are used to code the quantized spectral values. A modern audio coder, which works for example according to the MPEG-2 AAC standard, uses various Huffman code tables for coding the quantized spectral values, which are assigned to the spectrum in sections according to certain criteria. 2 or 4 spectral values are always coded together in one code word.

A difference between the MPEG-2 AAC method and the MPEG-Layer 3 method is that different scale factor bands, i. H. Different spectral values can be grouped into any number of spectral sections or "sections". In AAC, a spectral section or section includes at least four spectral values, but preferably more than four spectral values. The entire frequency range of the spectral values is therefore divided into adjacent sections, with one section representing a frequency band, such that all sections together comprise the entire frequency range which is covered by the spectral values after the transformation thereof.

A section is now assigned a so-called Huffman table from a plurality of such tables, as is the case with the MPEG Layer 3 method, in order to achieve maximum redundancy reduction. The Huffman code words for the spectral values in the ascending frequency sequence are now in the bit stream of the AAC method, which usually has 1024 spectral values. The information about the table used in each frequency section is transmitted in the page information. In addition to the spectral values, the scale factors of the MPEG-2-AAC (ISO / IEC JTC1 / SC29 / WG11 IS 13818.7) are also subjected to Huffman coding in order to further reduce the amount of bits to be transmitted. To further increase the efficiency, the scale factors are difference-coded within a frame, ie within an associated windowed block of samples, which is transformed into the frequency range. The difference is determined on the basis of a starting value, typically the first scale factor of a frame, which is absolutely given. This is particularly efficient for compression because small changes from one scale factor to the next are very likely.

A disadvantage of the Huffman code used is the fact that it has practically no redundancy. Although this is absolutely desirable for reasons of bit saving or data compression, it nevertheless leads to the fact that no redundancy is available in order to be able to maintain an error resilience.

If a Huffman-coded signal is transmitted over a faulty channel, there is almost no possibility in the decoder to "save" any valid values after the error has occurred. This is briefly explained using the Huffman-coded scale factors. As previously mentioned, the Huffman code is a variable length code. This is precisely the essence of the Huffman code, such that very frequently occurring values are assigned very short code words, while values that occur less frequently are given rather long or very long code words. In the bit stream syntax of the audio encoder addressed, the Huffman code words are written in succession for one frame in the bit stream. The beginning of the code word for a scale factor can only be determined if the corresponding preceding code word is correctly recognized, i. H. has been decoded.

For example, this means that between 40 and 100 scale factors occur in one frame, depending on how many scale len factor bands have been generated. This also means that about 40 to about 100 scale factors per band are Huffman encoded. The code words for the individual scale factors are simply written one after the other in ascending order in the bit stream. If the transmission of the bit stream over a faulty channel, such as a radio channel, leads to a bit error that changes the length of the code word that is assigned to the very first scale factor, it is impossible to decode the scale factors of the entire frame without errors , since the decoder has no way of determining the start of the code word for the second scale factor. Therefore, although under certain circumstances all other scale factors apart from the scale factor at the beginning, which was disturbed in the example, have been correctly transmitted, there is no longer any possibility in the encoder to decode the correctly transmitted scale factors.

The U.S. Patent No. 5,488,616 A is concerned with a system for supplying reversible codes of variable length. For this purpose, an asymmetrical reversible code is generated from a non-reversible code of variable length, which is only generated provisionally. The non-reversible variable length code is also converted to a symmetric reversible code. A selection device selects either the asymmetrical reversible code or the symmetrical reversible code as the output signal. The symmetric reversible code is represented by a complete code tree, in which all branches are terminated either by symmetrical code words or by branch points, these branch points again being terminated by a symmetrical code word or leading to further branch points. The code tree therefore contains only valid, i.e. H. symmetrical, code words.

The specialist publication "A tool for generating bit error resilient VLC tables" by Göran Bang and by Göran Roth, proposal for ISO / IEC JTC 1 / SC29 / WG11 dated July 1996 outlines a concept for encoding and decoding video information and audio information using a variable length code (VLC) suitable for use in connection with error prone channels. When a bit error is detected in the forward direction of a received bit stream, decoding is carried out in the reverse direction. If a bit error is also detected in the reverse direction decoding, the reverse direction decoding is also canceled. The code used is an unbalanced, fixed-length code into which a symmetrical, variable-length code is mixed such that a certain number of bits of a fixed-length code word is followed by a bit of a symmetrical, variable-length code word. The symmetrical code words of variable length serve only for the robustness of the error and do not carry any useful information. On the receiver side, the symmetrical code words of variable length are first extracted and analyzed for transmission errors.

A disadvantage of this mixed code is the fact that errors which occur in the fixed length code words cannot be determined, since only the symmetrical variable length code words are examined. On the other hand, undisturbed code words of fixed length can be identified as faulty if the associated code words of variable length are disturbed.

The object of the present invention is to provide a concept for entropy-coding information words and for decoding entropy-coded information words, which enables better error detection when the entropy-coded information words are transmitted over a faulty channel.

This object is achieved by a device for entropy coding according to claim 1 or 36, by a device for decoding entropy-coded information words according to claim 10, by a method for entropy coding according to Claim 24 and solved by a method for decoding according to claim 33.

The present invention is based on the finding that only the information words can be effectively transmitted in an error-resistant manner, which can be achieved by reversible, e.g. B. symmetrical, codewords are encoded. Only reversible code words allow forward and backward decoding of a sequence of code words that is uniquely assigned to a sequence of information words. In contrast to the Huffman code, which has asymmetrical code words, but which is almost optimal for data compression reasons, a symmetrical code has a higher redundancy. This redundancy can advantageously be used for error detection. However, in order not to sacrifice too much compression gain in the sense of error protection, according to the present invention not all information words are encoded by means of symmetrical code words, but rather only the information words which lie in a certain range of information words. Information words that are out of range are not encoded using the symmetric code, but can be Huffman encoded according to a preferred embodiment of the present invention. A compromise is thus made between robustness on the one hand and data compression on the other.

Another important aspect for the size of the range of information words encoded by symmetrical code words is the fact that a short code, i.e. H. a small code table, is desirable. The size of the area implicitly defines the length of the longest code word, since with an increasing number of code words in the table, the length of the valid code words also increases.

The error limitation is carried out according to the invention in that a decoder recognizes invalid, ie non-reversible, codewords and concludes that a transmission There is a supply error because such a code word was not generated in the encoder by definition. The probability that a disturbance leads to an invalid code word is greatest when there are only a small number of code words. If there is a very large number of code words, the probability that a fault leads to an invalid code word becomes smaller and smaller since the length of the invalid code words also increases more and more.

The method according to the invention is particularly advantageous where the information words to be coded are essentially within a range and information words are only outside the range with a low probability. The smaller the range, the less symmetrical code words are required and the better the error detection, which could be increased by adding artificial invalid code words. An attempt is therefore made to keep the area of the information words which are coded by means of symmetrical code words as small as possible in the sense of an effective error limitation, and yet to be so large that the information words are very likely to be within this area and are encoded symmetrically in order to to create sufficient overall robustness for errors.

A preferred application of the present invention is entropy coding of scale factors of a transformation-coded audio signal, since in this application 98% of the scale factor values occurring lie statistically within a manageable range which can be coded by symmetrical code words which are not yet excessive Have length. If an information word entropy is to be encoded that lies outside this range, an additional value is transmitted, which is called "escape". The escape value is preferably Huffman-coded and transmitted separately from the symmetrically coded scale factors in the audio bit stream. The purpose of the escape coding according to the invention is therefore to be able to cover a large range of code words despite a relatively small RVLC table with good error detection properties. The coding efficiency hardly suffers in said preferred application, since escape-coded values occur only rarely here.

The application of the present invention to the scale factors of a transform-coded audio signal is particularly advantageous because even minor disturbances in the scale factors due to a non-ideal channel lead to highly audible disturbances, since it is known that a scale factor weights several spectral lines ultiplicatively. Furthermore, since the scale factors only make up a relatively small part of the total bit quantity compared to the coded spectral values, protection of the scale factors by means of a redundant code does not lead to any significant additional expenditure of bits. This low additional effort is more than justified by the error-proofness of the scale factors, which can introduce disruptive interference into an audio signal in comparison to their bit quantity.

However, the present invention is not limited to the entropy coding or decoding of scale factors, but is advantageous wherever information words are to be coded which are very likely to be in a range such that without relatively great loss of efficiency with relatively short symmetrical code words can be used, values outside the range can be encoded by escape sequences.

Preferred embodiments of the present invention are explained below with reference to the accompanying drawings. Show it:

1 shows a schematic block diagram for an encoder according to the invention; 2 shows a schematic block diagram for a decoder according to the invention;

Fig. 3 is a schematic representation of the stream of code words processed by the decoder shown in Fig. 2;

4A to 4C symmetrical codes according to the prior art; and

5 shows a reversible code according to the present invention.

Before going into the figure in detail, however, some general aspects of entropy coding are presented. In particular, the special features of the coding of differentially coded values are dealt with, which can advantageously be combined with coding by means of symmetrical code words and escape values.

The code according to the invention represents an entropy code which, like the frequently used Huffman code, assigns short code words to frequently occurring values and long code words to rarely occurring values. However, the code according to the invention differs from the Huffman code as follows:

First of all, in contrast to the Huffman coding, the coding used allows decoding from both sides (forward and backward). This is also referred to in technology as "Reversible Variable Length Coding" (RVLC). Backward decoding is possible if reversible, e.g. B. symmetric code words, d. H. if a code book or a code table with reversible code words is used.

Furthermore, the use of an RVL code allows the effect that the code table, ie the set of available code words, does not form a "complete tree". It there are code symbol sequences that do not result in a valid code word, ie result in code words that are not symmetrical. The gaps that arise in the tree can be used to detect transmission errors since they indicate that a received code word has never been generated by a transmitter.

Since Huffman coding is optimal in terms of information theory, it makes no sense to use a different code if the only goal is to maximize data compression. However, the Huffman code is not advantageous when it comes to increasing error sensitivity with as little loss of compression efficiency as possible. By coding with code words of variable length, an error can propagate from the disturbed code word to all subsequent code words. A bit error thus falsifies the length of a code word, which means that all subsequent code words in a code word sequence can no longer be decoded, since the decoder has no knowledge of where the corresponding code words begin.

Instead of the variable length code, e.g. B. the Huffman code, now uses a reversible code of variable length, an error that changes the length of a code word can usually be diagnosed very quickly due to the presence of invalid code words. If an invalid code word becomes known, the decoding is aborted. This is not possible with Huffman coding because all code words are valid and therefore no invalid code words exist. In addition, the RVLC allows decoding from behind, which makes it easier to isolate the error. The following example shows this. A code table according to the invention could, for example, be as follows:

Information word codeword 3 110011

2 11011

1 111

0 0

-1 101

-2 1001

-3 10001

The range for information words that can be encoded by means of symmetrical code words using this code table thus ranges from -2 to +2. Values outside these limits, i. H. which are greater than 2 in amount, are given an escape value in addition to the symmetrical code word for -3 or +3.

Code words that cannot appear in the code table shown are the following:

10000 11000 110010 11010

With regard to a more detailed representation of a code according to the invention, reference is made to FIG. 5 and the corresponding discussion further below.

In the following, a numerical sequence 1, -3, 1, 2, -1 is to be considered as a fictitious example, which is to be transmitted via a faulty channel:

Table 2

Sequence of information words: 1, -3, 1, 2, -1 Sequence of code words: 111 10001 111 11011 101 If the case is considered that the twelfth bit is disturbed by an error generated by the channel, the following bit sequence results:

disturbed sequence of code words: 111, 10001, 111, 01011, 101

Decoding five values from the front gives 111, 10001, 111, 0, 101, i.e. H. 1, -3, 1, 0, -1.

However, the decoding from behind gives the following sequence:

101, 11010,

d. H. only -1 and an invalid code word. From this simple example it can be seen that by decoding from the back, i. H. due to the backward decoding, the invalid code word 01011 is recognized very quickly. Furthermore, the error can be very quickly localized and identified by aborting the decoding process after the invalid code word. Reverse decoding therefore reports an error in the range from the last bit to the eighth bit from behind.

The two decoded series of numbers are as follows. A bold print in the table below means that the values can be incorrect:

forward: 1 -3 1 0 -1 backward: x x x x -1

The extent to which a fault can be narrowed down depends on the type of fault and on the fault concealment technology implemented. Known error concealment techniques consist of simply replacing an incorrect value with its neighboring intact value. on the other hand If both intact values bordering on an error are known, weighted average values from the left and right margins can be used to artificially replace the incorrect value, ie to obscure them. Still other error concealment techniques use interpolation using two adjacent values between which there is an error. Likewise, one-sided prediction can be taken from the front or from the back to replace an incorrect value with a "possibly relatively intact" value.

1 shows an entropy encoder according to a preferred embodiment of the invention. Information words to be coded are fed into a decision maker 10. For the sake of simplicity, information words in the tables above consisted only of integers. In an audio encoder, information words that are to be entropy-encoded, such as B. scale factors, for example as eight-bit values. The expression "information words" is thus intended to encompass any type of representation by which information to be coded can be represented.

It is determined in the decision maker 10 whether the information word is in a range of information words or outside the range of information words. The range is determined by the code table implemented in the encoder. If it is determined in the decision maker that an information word to be encoded lies within the range, the same is transmitted to a device 12 for assigning a symmetrical code word from a group of symmetrical code words, ie from the code table, in such a way that a symmetrical code word is assigned to the information word becomes. On the other hand, if the decision maker 10 decides that the information word lies outside the area determined by the code table, this information word is transmitted by the decision maker 10 to a device 14 for generating an additional value, such that the device 14 in a preferred exemplary embodiment of the present the invention determined the escape value. The device 14 principally comprises two outputs, ie an output for writing the escape value into an escape area of the bit stream, and on the other hand an output which is connected to a bit stream formatter 16 which has a stream of code words or a sequence of Generates code words which is assigned to the sequence of information words fed in at the input of the decision maker 10.

For a more detailed explanation of the functioning of the device for generating an additional value or escape value, which is identified in FIG. 1 by reference numeral 14, refer to FIG. 3. Fig. 3 shows a continuous stream 30 of "potentially" symmetrical code words, the term "potentially" is intended to indicate that the stream 30 is already over a non-ideal channel, e.g. B. a radio link has been transmitted, which may have caused bit interference. The stream consists of individual symmetrical code words 30a, 30b, 30c, etc., all of which are within the range defined by the code table comprising symmetrical code words. However, the stream of potentially symmetrical code words also includes symmetrical code words 31, 32 which stand for information words at the edge of the range. The code words 30a-30c are generated by the assignment device 12 and fed into the bitstream formatter 16. In a preferred exemplary embodiment of the present invention, the code words which exist for information words at the edge of the region are generated by the device 14 and from there are fed into the bit stream formatter 16, which forms the stream 30 shown in FIG. 3. Codewords 30a-30c and 31 and 32 represent information words from -7 to +7, i. H. Information words to which symmetrical code words are assigned. If the information word to be coded has a value of, for example, +12, the sum of the symmetrical code word 31 and the escape value gives the value +12.

A decoder that encounters code word 31 thus recognizes continues that this is a code word at the edge of the area, which is why the decoder for decoding the information word "goes" into the escape area by means of the link A in order to determine that there is an escape value of 5 in the present example is. The device 14 for generating thus performs two functions in accordance with a preferred exemplary embodiment of the present invention. First, it supplies the code word for the edge of the area in the stream 30 from symmetrical code words. On the other hand, it forms the difference between the information word to be coded and the code word at the edge of the area and generates an escape value which represents the difference. Of course, the escape value can again be entropy-encoded using the coding method according to the invention. For data compression reasons, however, it is preferred to encode the escape value using a Huffman code. From Fig. 3 it can also be seen that the escape value is not written into the stream of symmetrical code words, but at another location in the bit stream.

If the value -12 is now to be coded, the decision maker 10 will determine that this value lies outside the range defined by the code table with symmetrical code words. The device 14 for generating an additional value will therefore on the one hand output the code word for the value -7 to the bit stream formatter 16 and on the other hand the difference, i. H. 5, write in the escape area. The value -12 then results from the combination of the value -7, for example the code word 32 in FIG. 3, and the escape value 34 via the link represented by the arrow B.

A value of +7 would be in the preferred embodiment of the present invention as a code word for +7, i.e. H. as the code word 31, and 0 in the escape area, i.e. H. as the escape value 33.

In deviation from the described embodiment, it is it is not essential that the device 14 for generating an additional value forms the difference between the information word to be coded and the information word at the edge of the area and, on the one hand, writes a symmetrical code word in the stream 30 of symmetrical code words and, on the other hand, writes the difference in the escape area. It is alternatively also possible that the entire information word is written into the escape area, and in the stream of symmetrical code words only a placeholder, a specific bit combination or the like is inserted either by the device 14 or by the bit stream formatter 16 in order to be connected downstream To signal the decoder that it must go into the escape area at this point in the bit stream. However, the method illustrated has the advantage that at least the part of the information word that lies in the area covered by symmetrical code words is coded using a symmetrical code word, whereby only the difference, which is coded using a Huffman code, for example, less secure or less error-prone. On the other hand, alternative driving would have the advantage that no addition or difference formation has to be carried out and that the less redundant code is used for an information word outside the range. However, it is disadvantageous that the information word lying outside the range cannot then be decoded backwards. However, either methods known in the prior art or the coding method according to the invention can be used to protect the escape area in order to create secure conditions here as well.

2 shows a preferred exemplary embodiment of a decoder according to the present invention. A sequence of code words or a stream of “potentially symmetrical code words” 30 is fed into a memory 20, which an analyzer 21 can access to analyze a sequence stored in the memory 20. On the one hand, the analyzer 21 comprises a device for detecting a symmetrical code words from the sequence 30 of code words and, on the other hand, means for detecting a predetermined code in the sequence 30 of code words. If the analyzer 21 detects an intact symmetrical code word, it transmits this to a device 22 for assigning a specific information word to the code word based on a previously known code table which must correspond to the code table used in the encoder (FIG. 1). However, if the analyzer 21 detects a predetermined code, which in the present example is the code word for an edge of the area, it sends this code word to a device 23 for determining an additional information word outside the area. In the preferred exemplary embodiment, the device 23 will access the escape area when a code word 31 occurs in the stream 30 and retrieve the corresponding escape value there and add or subtract it to the information word that corresponds to the code word 31.

The assignment of a predetermined code, in the exemplary embodiment of the code word for an information word on the edge, to a code word in the escape area can be done in various ways. The simplest is to use a continuous pointer, with both the escape area and stream 30 being synchronized. This synchronization is provided in the preferred embodiment of the present invention, in which scale factors are entropy-coded, by always processing a block or frame of an audio signal. For each block, the escape area and stream 30 of symmetric code words are initialized so that the use of a continuous pointer gives correct results.

The encoder according to the invention further comprises means 24 for reversing the order of the decoder, which is activated by the analyzer 21. If the analyzer 21 finds an asymmetrical code word in the stream 30, it will activate the device 24 for reversing the order, since an asymmetrical code word in the stream 30 potentially has symmetric code words cannot occur. The means 22 for assigning and the means 23 for determining then work in the reverse order from the other end of the sequence of code words in order to limit the error with a decoding from the rear, in such a way that as few values as possible have to be replaced by an error concealment.

In reality it can happen that an entropy decoder does not immediately recognize a disturbed code word, since the disturbance did not lead to an invalid code word. It therefore decodes beyond the error until it encounters an invalid code word as a result of a subsequent error and the coding then stops. The backward decoder may then also decode from the other end, possibly also beyond the disturbed code word, and terminate at some point if the code word is invalid. This creates an overlap area in which both the entropy forward and the entropy backward decoders have provided output values. The error is thus limited to the overlap area and it can be determined that the decoded values outside the overlap area are correct.

If differential coding has been carried out in the encoder, the corresponding decoder furthermore comprises a differential decoder 25 which cancels the differential coding generated in the encoder. According to the invention, the difference decoder 25 is also activated by the device for reversing the sequence 24 in order to carry out a backward difference decoding in order to also use reverse-coded information words transmitted by the devices 22 and 23 to the device 25 to produce completely decoded information words again. It should be noted that the reverse differential decoder and the forward differential decoder may be separate devices, or may be implemented by a single device, with an addition in forward differential decoding is performed while subtraction is being performed in reverse differential decoding.

The combination of a differential coding with the coding method according to the present invention is particularly advantageous, since only through a differential coding with a suitably chosen output value of the difference formation does the absolute information words, e.g. B. "shifted" by a symmetrical area to zero.

However, in order for differential decoding 25 in the reverse direction to be possible from the other end of an information word sequence, an additional value must be added at the end of the sequence of information words in the encoder, such that the differential decoder knows from where a difference Decoding should be started from behind. If a defined start value has been used in the difference coding, an additional value is an additional value at the end of the sequence of difference-coded information words, which indicates the difference from the last information word to the defined or predetermined start value. The difference is of course also entropy-encoded and is preferably entropy-encoded using a symmetrical information word, such that this value is well protected to allow reverse decoding. If the first information word of a sequence of information words is taken as the starting value for the differential coding in the encoder, then it makes sense to add the absolute value of the last information word as an additional value at the end of the sequence. This last value will then almost certainly not be in the range of information words which are coded with symmetrical code words.

As has already been mentioned, a preferred application of the present invention is in the coding of scale factors, which were previously differentially coded and then Huffman coded. In the prior art, one Huffman table with 121 code words used to code values in the range from -60 to +60. Since the number of scale factors to be encoded is very small compared to the number of spectral values, a typical value is 40, a relatively "fast" detection of errors is absolutely necessary, such that the decoder aborts fewer values after decoding, so that a Errors can be localized relatively well. A "small" code table is therefore used, which means that the number of symmetrical RVLC code words should be small.

The smaller the code table or code book, the more likely it is that errors will be detected earlier. Therefore, the range for information words that are encoded symmetrically runs from -7 to +7. The escape value is transmitted for values in the range from 7 to 60. Preferably the same Huffman is encoded. The "escape" table thus consists of 54 entries for values between 0 and 53. Every time the receiver decodes a -7 or a +7, it must decode and add or subtract the associated escape value.

Statistically speaking, the interval -7 to +7 covers 98% of the scale factor values that occur, such that escape values are not particularly common. If more frequent escape values occur, or if even greater attention is paid to error safety, various known methods and the method according to the invention can be used in order to make the escape values more error-proof.

To explain the reversible variable-length code according to the invention, reference is first made to FIG. 4A, in which a known symmetrical code is shown, which is disclosed, for example, in the above-mentioned specialist publication by Göran Bang and Göran Roth. This code is defined by a code tree that has a root 40 and a branch point 42. Since it is a binary code, two branches 43, 44 are at the root. brings, the branch 43 connects the root 40 with an end point that defines the valid code word "1". The branch 44 connects the root 40 to the branching point 42, from which two branches 45, 46 lead away. Branch 46 is connected to an endpoint which defines the second valid code word of this code "00", while branch 45 defines an invalid code word of this code, ie "01". The code word "01" is invalid because it is asymmetrical. With regard to the further notation, it should be pointed out that invalid code words are framed in FIGS. 4A and 4C and in FIG. 5. The code shown in Fig. 4A thus comprises only two valid code words, ie "1" and "00", and only one invalid code word, ie "01", which is the same length as the second valid code word "00".

A slightly longer code is shown in Fig. 4B. In contrast to FIG. 4A, the code in FIG. 4B contains a further valid code word "010" and also an invalid code word which, like the additional valid code word, is now 3 bits long and has the value "011". In contrast to FIG. 4A, branch 45 is not connected to an end point, but rather to a further branch point 47, from which two branches start, the first outgoing branch reaching for the additional valid code word "010", while the other branch is invalid for the only one Code word "011" is sufficient.

FIG. 4C is the logical continuation of FIGS. 4A and 4B, since branch point 47 is now connected to a further branch point 48, from which in turn two branches start, one branch defining an additional symmetrical code word "0110", while the end point of the other branch defines the only invalid code word "Olli" that has the same length (4 bits) as the longest code word of the code tree, i. H. the code table.

It can also be seen from FIG. 4C that no valid code words of the same length exist. This also applies to the codes in Figures 4A and 4B. The codes shown in FIGS. 4A to 4C are used in the specialist publication by Göran Bang and Göran Roth only as a security grid and not as a code for coding information, since such a code, as is readily achieved by logical continuations of the data in FIGS. shown codes can be seen, with a sufficiently high number of code words becomes very long. In addition, the robustness of errors corresponding to longer codes is very low, since there is always only one invalid code word and, moreover, this invalid code word is as long as the longest valid code word. Coding information with the known codes is therefore not appropriate since the code words become very long if an appropriate range of information values is to be coded, and because there is always only one invalid code word which is also very long. A decoder will therefore not immediately detect an error and commit a large number of subsequent errors before it encounters an invalid code word and aborts the decoding. So the error is narrowed down badly.

The code table according to the invention shown in FIG. 5 overcomes these disadvantages in that there is at least one branch point in the code tree from which two branches start, both of which are connected to a branch point instead of an end point. In the preferred exemplary embodiment for the reversible code with variable length shown in FIG. 5, these are branch points 50a, 50b, 50c and 50d. 5 further comprises a root 52 from which two branches 53 and 54 originate, branch 53 being connected to an end point which defines the first and shortest code word 0, which in the preferred exemplary embodiment has the information value " 0 "is assigned. The code word 0 has the shortest length and is therefore associated with the information value that occurs most frequently in the sense of entropy coding. In the preferred embodiment of the present invention, in which the information values are differently encoded, it has been found that the value 0 is most likely to occur, particularly in the differential coding of scale factors.

The other branch, starting from the root 52, i.e. H. branch 54, according to the invention, does not end in a code word with the length of 2 bits but leads to branch point 50a, which in turn is connected to further connection points 57, 58 via two branches 55, 56. The connection points 57 and 58 are in turn provided with corresponding branches 59, 60 with end points which define the valid code words 101 and 111. In contrast to the prior art, it can be seen here that by dispensing with a code word with a length of 2 bits, 2 code words can be obtained which are of the same length, i. H. 3 bits in FIG. 5. These are code words 101 and 111. In the code table shown in FIG. 5, they are assigned the information values "-1" and "1". From the point of view of entropy coding, it is expedient to provide code words of the same length for two information values which are very likely to occur with the same frequency.

It can be seen from FIG. 5 that two branches 62, 63 start from the connection point 50c, the branch 63 being connected to a connection point 64 which is connected via branches to a valid code word 110011 and to an invalid code word 110010. If a decoder encounters the invalid code word 110010, for example, it will abort the coding process since such a code word is irreversible and has never been generated in an encoder.

The formation and assignment of further code words to information values can be seen in FIG. 5. However, another special feature should be pointed out. A preferred application of the code according to the invention shown in FIG. 5 is the use of the reversible code in connection with escape values. As has already been shown, an information value outside the range from "-7" to "+7" is replaced by the code word for the corresponding information value at the edge of the range and further difference encoded in an escape table. Therefore, there is a higher probability that the value at the edge of the range, ie "-7" or "7" must be encoded in the encoded bit stream. Furthermore, there is an equal probability that "-7" or "+7" will occur. According to a preferred embodiment of the present invention, the information values "-7" and "+7" are encoded by code words of the same length, ie 1000001 and 1100011, respectively, these code words being at the same time shorter than the longest occurring code words, which in this case are the code words for "-6" and "+6" to create a code table that is as good as possible from an entropy point of view.

It can also be seen from FIG. 5 that eight invalid code words 66a-66h exist, whereas a reversible code according to the prior art only ever has or can have one code word. A large number of invalid code words and particularly relatively short invalid code words, e.g. the code words 66e, 66f provide a high level of error resilience, such that a decoder aborts the decoding as quickly as possible after a disturbed information value, in such a way that an error can be narrowed down as narrowly as possible.

Finally, it should be summarized that the reversible, variable-length code shown in FIG. 5 is particularly well suited for error-resistant entropy coding of information values, since on the one hand there are a relatively large number of relatively short invalid code words, and on the other hand for the price of a shorter code word (in the example the code word "11") two code words (in the example 101 and 111) are obtained, which are longer but equally likely. Although the omission of a valid short code word in the sense of entropy coding can actually be avoided, the same delivers according to the invention in applications in which error-resistant entropy coding is to be carried out and in which two information values are also included a relatively high probability and particularly with approximately the same probability of occurring, a good solution. With regard to the code table representation in tree form, this waiver is accomplished by the fact that special branch points exist, from which two branches originate, but which are both connected to further branch points instead of one end point.

Claims

claims

1. Device for the entropy coding of information words, with the following features:

means (12) for assigning a reversible code word from a group of reversible code words to an information word which is in a range of information words, the group of reversible code words being designed such that a separate reversible code word is provided for each information word in the range is; and

means (14) for generating an additional value (33, 34) for an information word which is outside the range of information words.

2. The apparatus of claim 1, wherein the reversible code words are symmetrical code words.

3. Apparatus according to claim 1 or 2, wherein the means (14) for generating is arranged such that it also generates an additional value for an information word which lies on a limit of the range of information words.

4. Device according to one of the preceding claims, which is designed such that information words occurring with a high probability lie in the area.

5. Device according to one of the preceding claims, wherein the means for assigning is arranged such that the size of the area is determined depending on the information words to be coded such that a disturbance of a code word with a high probability that this code word by the Disturbance becomes an irreversible code word.

Apparatus according to one of the preceding claims, in which the device (14) for generating an additional value is arranged such that it forms, as an additional value, the difference between the information word and the information word on a next boundary of the area.

Device according to one of the preceding claims, further comprising:

means for assigning a code word from a further group of code words to the additional value (33, 34).

Apparatus according to claim 7, which is arranged such that, for entropy coding an information word lying on the boundary of the area or outside the area, the means (12) for assigning this information word assigns the code word which corresponds to the information word on the next boundary corresponds to the range; and

the means (14) for generating generates the difference between the information word to be encoded and the information word on the boundary of the area, the sign of the difference being determined by the sign of the information word on the boundary of the area, such that the information word to be encoded by the reversible code word for the information word on the boundary of the range and an unsigned difference is represented as an additional value (33, 34).

Device according to one of the preceding claims, further comprising:

a bit stream formatter (16) for writing the reversible code words in a first area (30) of the bit stream and for writing the additional values in a second area (33, 34) of the bit stream.

10. Device according to one of the preceding claims, further comprising the following feature:

a device for difference coding of information, which generates difference-coded information words on the basis of a starting value, successive difference-coded information words representing a difference value sequence.

11. The apparatus of claim 10, wherein the means for differential coding further comprises:

means for adding an additional element to the end of the difference value sequence, the additional element being determined in such a way that a backward decoding of the difference value sequence can be carried out.

12. Device for decoding entropy-coded information words using reversible code words, wherein reversible code words are assigned to an area of the information words, while an information word outside the area is represented by an additional value, having the following features:

means (21) for detecting a reversible code word (30a, 30b, 30c) from a sequence (30) of code words;

means (22) for assigning a particular information word to the detected code word based on a code table; means (21) for detecting a predetermined code (31, 32) in the sequence (30) of code words; and

means (23) for determining an additional information word (33, 34) outside the range on the basis of the predetermined code (31, 32).

13. The apparatus of claim 12, wherein the reversible code words are symmetrical code words.

14. The apparatus of claim 12 or 13, which is arranged such that the additional element (33, 34) is a non-symmetrical code word.

15. The apparatus of claim 14, wherein the non-symmetric code word is a Huffman code word.

16. The device according to one of claims 13 to 15, wherein the means (21) for detecting a reversible code word is arranged such that a non-reversible code word in the code word sequence (30) can be determined.

17. The apparatus of claim 16, further comprising:

means (24) for reversing an order in which the means (21) for detecting processes the code word sequence (30), the means (24) for reversing being responsive to the detection of a non-reversible code word.

18. Device according to one of claims 13 to 17, in which the information words are difference-coded, further comprising:

means (25) for differential decoding of the dif- reference-coded information words.

19. Apparatus according to claim 17 or 18, in which the sequence of information words is difference-coded starting from a start value, the sequence further having at its other end an additional value which is chosen such that a backward difference Decoding is feasible from the other end, the differential decoding means (25) being arranged to perform differential decoding from the other end in response to the reverse order means.

20. Device according to one of claims 13 to 19, wherein the predetermined code (31, 32) is the code word which is assigned to an information word on a boundary of the range of information words.

21. Device according to one of claims 13 to 20, wherein the predetermined code (31, 32) is a reversible code.

22. The apparatus of claim 20 or 21, wherein the means (23) for determining is arranged such that it forms the sum of the information word on a boundary of the area and the additional information word outside the area to obtain an information word that is outside of the range.

23. Device according to one of the preceding claims, in which the information words are scale factors of a transform-coded audio signal.

24. A method for entropy coding information words, comprising the following steps:

Assigning (12) a reversible code word from a group of reversible code words to an information onsword, which is in a range of information words, the group of reversible code words being designed such that a separate reversible code word is provided for each information word in the range; and

Generating (14) an additional value (33, 34) for an information word which is outside the range of information words.

25. The method of claim 24, wherein the reversible code words are symmetrical code words.

26. The method of claim 24 or 25, wherein the area comprises the information words occurring with a high probability.

27. The method according to any one of claims 24 to 26, wherein the step of generating (14) an additional value comprises the following step:

Forming a difference between the information word and the information word on a next boundary of the area as an additional value.

28. The method according to any one of claims 24 to 27, further comprising the step of:

Assigning a code word from a further group of code words to the additional value (33, 34).

29. The method of claim 28, wherein the assigning step comprises the following step:

Entropy encoding an information word that is on or outside the area boundary, assigning (12) the code word to that information word that is the information word on the next boundary corresponds to the range; and

in which the step of creating comprises the following step:

Generating (14) the difference between the information word to be encoded and the information word on the boundary of the area, the sign of the difference being determined by the sign of the information word on the boundary of the area, such that the information word to be encoded is by the reversible code word for the information word on the boundary of the range and an unsigned difference is shown as an additional value (33, 34).

30. The method according to any one of claims 24 to 29, further comprising the step of:

Writing the reversible code words in a first area (30) of the bit stream and writing the additional values in a second area (33, 34) of the bit stream.

31. The method according to any one of claims 24 to 30, further comprising the step of:

Differential coding of information, whereby difference-coded information words are generated starting from a starting value, successive difference-coded information words representing a difference value sequence.

32. The method of claim 31, further comprising the step of:

Adding an additional element to the end of the difference value sequence, the additional element being determined such that a reverse decoding of the Differential value sequence is feasible.

33. A method for decoding entropy-coded information words using reversible code words, wherein reversible code words are assigned to an area of the information words, while an information word outside the area is represented by an additional value, with the following steps:

Detecting (21) a reversible code word (30a, 30b, 30c) from a sequence (30) of code words;

Assigning (22) a particular information word to the code word based on a code table;

Detecting (21) a predetermined code (31, 32) in the sequence (30) of code words; and

Determining (23) an additional information word (33, 34) outside the area on the basis of the predetermined code (31, 32).

34. The method of claim 33, further comprising the steps of:

Reversing (24) an order in which the means (21) for detecting processes the code word sequence (30) in response to a non-reversible code word.

35. The method as claimed in one of claims 24 to 35, in which the information words are scale factors of a transformation-coded audio signal.

36. Device for entropy coding of information values, with the following features:

a device (12) for assigning a code word from a code table with a plurality of mutually different reversible code words of variable length to an information value, the code words of the code table being defined by a code tree (FIG. 5) which has the following features:

a root (52);

a plurality of branches (53, 54, 55, 56, 59, 60, 62, 63);

a plurality of branch points (50 a - 50 d, 57, 58), one branch leading to each branch point and two branches starting from each branch point; and

a plurality of end points of branches which define valid code words or invalid code words (66 a - 66 h), the valid code words being reversible and the invalid code words not being reversible,

two branches (53, 54) originating from the root (52), and

two branches (55, 56) starting from at least one branch point (50a, 50d) of the code tree, both branches being connected to a branch point (57, 58) instead of to an end point of a branch.

37. The apparatus of claim 36, wherein a branch (53) connected to the root (52) is also connected to a first order end point, the first order end point defining the shortest valid code word that corresponds to the information value that is most likely assigned.

38. The apparatus of claim 37, wherein the branching point (50a), which is connected to the other branch (54), which starts from the root (52), is connected to two further branch points (57, 58).

39. Apparatus according to claim 38, in which one of the two branches (59, 60), which are connected to the two further branch points (57, 58), are each connected to a second-order end point, such that two valid ones have the same length Code words are defined, the length of which is three times the length of the shortest code word, and each of which is associated with information values that are most likely to occur and are approximately equally likely.

40. Device according to one of claims 36 to 39, in which there is at least one invalid code word (66e, 66f) which is shorter than the longest valid code word.

41. Device according to one of claims 36 to 40, wherein the valid code words are symmetrical.

42. Apparatus according to any one of claims 36 to 41, wherein the code words are binary words, the shortest code word is one bit long, and there are two second shortest code words that are three bits long.

43. Device according to one of claims ^ 6 to 42, in which more than one invalid code word (66a-66h) exists.

44. Device according to one of claims 36 to 43, in which the code table is defined as follows: Information word codeword

0 0

-1 101 -2 1001 -3 10001 -4 100001 -5 10000001 -6 100000001 -7 1000001 +1 111 +2 11011 +3 110011 +4 1101011 +5 11000011 +6 110101011 +7 1100011

45. Apparatus according to one of claims 36 to 44, in which the code table assigns information words in an area to code words, the code words which are assigned to information words on a boundary of the area being shorter than other code words in the code table.

46. The apparatus of claim 45, wherein the information values outside the area are entropy-encoded by the code words on a border of the area and by an additional escape value.